Deep-dive lectures
Seven undergraduate-level lectures that broaden the assessment knowledge beyond the FastAPI track — pipeline, architectures, loss functions, ablations, and the math behind modern damage-assessment models.
Project Overview · xBD + ATC-20
Master lecture: multi-architecture study on xBD with operational ATC-20 deployment. Figures, tables, equations, and the umbrella for every other module.
- xBD dataset
- Pipeline
- ATC-20
Vision Transformers (ViT)
Patches, self-attention, multi-head MSA, encoder block, variants (DeiT, Swin, BEiT), and how residuals defeat vanishing gradient.
- 12 sections
- Q/K/V math
- Architecture diagrams
Focal Loss, Dice Loss & Class Weights
Why cross-entropy fails on imbalanced disaster data, the (1−pₜ)^γ modulating factor, region-overlap losses, and combined Focal+Dice training.
- Class imbalance
- Worked numerics
- PyTorch code
Siamese Neural Networks
Shared encoders, difference blocks, skip connections, attention gates, deep supervision, and Siamese variants for pre/post change detection.
- Twin networks
- xBD pipeline
- Deep supervision
Encoder · ResNet-50
Stem, four bottleneck stages, residual shortcuts, the 7×7 deepest feature map, and the four skip-tap points that feed the U-Net decoder.
- Backbone
- Bottleneck zoom
- Skip-taps
Decoder · U-Net
Symmetric expanding path: transposed convolutions, concatenated skip connections, double-conv blocks, dropout, and the 1×1 softmax head that paints the per-pixel damage map.
- UpConv
- Skip-concat
- Pixel-accurate
Ablation Study
What an ablation study is, why we need it, and a chart-by-chart walkthrough of the four-architecture comparison on xBD — explained for undergraduates.
- What-if analysis
- Bar charts
- Per-class F1
Hyper-parameters
Learning rate, batch size, optimiser, weight decay, dropout, schedulers, and the exact settings used for the xBD project — with figures, equations, and a tuning workflow.
- η, B, AdamW
- Schedulers
- Tuning cheat-sheet
Suggested 3-week reading schedule
A paced plan that assumes ~45–60 minutes of focused reading per day. Each row pairs a lecture with a concrete "you can do this after" goal. Stretch it to 6 weeks if you also code along in the FastAPI track.
| Week | Days | Lecture | Time | By the end you can… |
|---|---|---|---|---|
| Week 1 | Day 1–2 | L0 · Project Overview | 90 min | Understand the end-to-end pipeline, the xBD dataset, and what an ATC-20 placard is. |
| Week 1 | Day 3–5 | L1 · Vision Transformers | 2 × 60 min | Master patches, self-attention, multi-head, and why residuals defeat vanishing gradients. |
| Week 2 | Day 1–2 | L2 · Focal, Dice & Class Weights | 75 min | Explain why plain cross-entropy fails on xBD and how each loss component fixes a different failure mode. |
| Week 2 | Day 3–4 | L3 · Siamese Networks | 75 min | Reproduce the shared-encoder + difference-block diagram from memory. |
| Week 2 | Day 5 | L4 · Encoder · ResNet-50 | 60 min | Walk the stem → 4 bottleneck stages → skip-taps without checking notes. |
| Week 3 | Day 1 | L5 · Decoder · U-Net | 60 min | Trace how a 16×16 deep feature becomes a 512×512 damage map via skip-concat + up-conv. |
| Week 3 | Day 2–3 | L6 · Ablation Study | 75 min | Rank the four components by Δ-F1 and justify the order in writing. |
| Week 3 | Day 4–5 | Recap + Quiz | 90 min | Score ≥ 80% on the 50-question quiz; revisit any lecture where you scored < 70% on its questions. |
- Read the "In plain English" primer at the top.
- Skim every figure caption before reading the body.
- Re-derive at least one equation on paper.
- Open the related quiz questions and answer from memory.
≈ 10 hours of reading + 3 hours of quiz/review across three weeks. Realistic for a student carrying a normal course load.
Read L0 → L2 → L3 → L6 in that order. You'll know the pipeline, why the losses matter, the change-detection trick, and which components carry the model's accuracy.